A custom MTPLX-forged version of Qwen/Qwen3.8-27B, optimized for Apple Silicon with a compact 16.924 GB complete artifact size (15.762 GiB) while retaining native MTPLX multi-token prediction support.

This build is intended for users who want a capable 27B-class Qwen model that leaves headroom for a practical context window and lightweight applications on machines with constrained unified memory.

Key details

Field	Value
Public model name	`Qwen3.8-27B-MTPLX-16.9GB-4bit`
Base model	`Qwen/Qwen3.8-27B`
Base revision	`1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`
Complete artifact size	16.924 GB / 15.762 GiB
Quantization	4-bit affine, group size 64
Measured body precision	4.501 bits/weight
MTPLX support	Native MTP sidecar, FP16, 15 tensors
Runtime compatibility	`verified-native` in MTPLX 2.9.0
Max MTP depth in artifact	3
Recommended tested depth	Depth 2 on Apple M2 Pro
Vision tower	Preserved (`model-vision.safetensors`)

Built with MTPLX

This model was forged and optimized using MTPLX.

MTPLX runs local LLMs on Apple Silicon with native multi-token prediction where supported by the model architecture. The MTP sidecar in this artifact is preserved as native FP16 weights rather than being requantized.

Forge recipe

json

{
  "body_bits": 4,
  "body_dtype": "auto",
  "body_group_size": 64,
  "body_mode": "affine",
  "mtp_policy": "keep_bf16"
}

--dtype auto on Apple M-series hardware maps non-quantized tensors, including the native MTP sidecar floats, to FP16.

Verification summary

mtplx inspect --require-mtp reports this artifact as:

support level: verified-native
runtime compatibility: native
unverified model: false
MTP tensor count: 15 / 15 expected
quantization: 4-bit affine, group size 64

Usage

Install MTPLX from the project instructions, then load this model by Hugging Face repo id or local path in MTPLX.

bash

mtplx inspect --require-mtp <repo-or-local-path>
mtplx start --model <repo-or-local-path>

For best results on similar Apple Silicon hardware, start with MTP depth 2 and tune locally:

mtplx tune --model <repo-or-local-path> --retune

Notes and limitations

This is an MLX/MTPLX-oriented artifact for Apple Silicon.
Runtime memory depends on context length, KV cache settings, server mode, and other resident applications; the listed 16.924 GB is the complete on-disk artifact size, not a universal maximum runtime allocation.
Model weights remain governed by the upstream base model license and terms, plus any licenses attached to the tooling used to create this artifact.

Attribution

Base model: Qwen/Qwen3.8-27B
MTPLX: https://github.com/youssofal/MTPLX

LloydAgent/Qwen3.8-27B-MTPLX-16.9GB-4bit

Qwen3.8-27B-MTPLX-16.9GB-4bit